RelationNetworkModel<T, TInput, TOutput>

Models & Types · AiDotNet.MetaLearning.Models

Relation Network model for few-shot classification.

Implements: IModel<TInput, TOutput, ModelMetadata<T>>

For Beginners

After the Relation Network sees the support examples (the few labeled examples for each class), this model remembers them and uses them to classify new query examples. It does this by computing how “related” the query is to each class.

How It Works

The adapted state of a Relation Network for one task: its own copy of the embedding module, the support set’s embeddings and labels, and the learned comparison. Predict returns each example’s relation score with every class, [rows, NumClasses], for Tensor and Matrix outputs - a class with no support example scores zero - and the class with the highest score for a Vector output.

Constructors

Constructor Summary
RelationNetworkModel(IFullModel<T, TInput, TOutput> featureEncoder, RelationModule<T> relationModule, TInput supportInputs, TOutput supportOutputs, RelationNetworkOptions<T, TInput, TOutput> options) Initializes a new instance of the RelationNetworkModel around one relation module.

Properties

Property Summary
ModelMetadata<T> Metadata { get; }

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
TOutput Predict(TInput input)
Vector<T> GetParameters()
void Train(TInput inputs, TOutput targets)
void UpdateParameters(Vector<T> parameters)